Exploiting Hardware Heterogeneity within the Same Instance Type of Amazon EC2

Transcription

1 Exploiting Hardware Heterogeneity within the Sae Instance Type of Aazon EC2 Zhonghong Ou, Hao Zhuang, Jukka K. Nurinen, Antti Ylä-Jääski, Pan Hui Aalto University, Finland; Deutsch Teleko Laboratories, Gerany Abstract Cloud coputing providers ight start with nearhoogeneous hardware environent. Over tie, the hoogeneous environent will ost likely evolve into heterogeneous one because of possible upgrades and replaceent of outdated hardware. In turn, the hardware heterogeneity will result into perforance variation. In this paper, we look into the hardware heterogeneity and the corresponding perforance variation within the sae instance type of Aazon Elastic Copute Cloud (Aazon EC2). Standard large instance is selected as the exaple. We find out that there exist three different subtypes of hardware configuration in the standard large instance. Through a set of detailed icro-benchark and application-level benchark easureents, we observe that the perforance variation within the sae sub-type of instance is relatively sall, whilst the variation between different sub-types can be up to 60%. By selecting better-perforing instances to coplete the sae task, end-users of Aazon EC2 platfor can achieve up to 30% cost saving. 1 Introduction Cloud coputing attracts a significant aount of attention fro industry, acadeia, and edia because of its on-deand, pay-as-you-go, etc, characteristics. As a representative and one of the ost widely adopted public cloud platfors, Aazon Elastic Copute Cloud (Aazon EC2) has been used for a host of sall and ediusized enterprises (SMEs) for various usages. Aazon EC2 was introduced in 2006, and supports a wide arrange of instance types. Naturally, these different types of instances are likely hosted by heterogeneous hardware. Over tie, because of hardware upgrade and replaceent, it would be interesting to investigate the following issues: (1) Does the sae type of instance utilize hoogeneous or heterogeneous hardware configuration? (2) If heterogeneous hardware is used, what is the resulting perforance variation? In this paper, we try to answer the aforeentioned two questions by utilizing the standard large instance type, i.e. 1.large. Siilar results are observed for the other types of instances within the sae standard faily, including sall (1.sall), and extra large (1.xlarge) instances. Our contributions are as follows: (1) We observe that within the sae instance type, Aazon EC2 uses heterogeneous hardware to host the instances. (2) The variation of the sae sub-type of instances, i.e. hosted by identical hardware, is relatively sall, whilst the variation aong different sub-types of instances, i.e. hosted by heterogeneous hardware, can reach up to 60%. (3) Copared with taking the rando instances assigned by Aazon EC2 platfor, by selecting betterperforing instances to coplete the sae task, EC2 users can acquire up to 30% of cost saving. The rest of the paper is structured as follows. In Section 2, we present background and related literature of Aazon EC2 study. Section 3 details the icrobenchark easureents and application-level bencharks. Section 4 analyzes the potential cost saving for EC2 end-users. In Section 5 we conclude the paper and present ideas for future work. 2 Related Work Several studies have been conducted to analyze various aspects of Aazon EC2. Garfinkel [4] conducted a easureent study of various Aazon Web Services (AWS) to evaluate the feasibility and cost of oving a largescale research application fro localized server to Aazon offering. Palankar et al. [8] perfored easureents focusing on Aazon S3 to testify its ability to provide stable storage support for large-scale scientific

2 coputation application. Walker [12] studied the perforance of Aazon EC2 high-perforance cluster copute instances against a locally configured equivalent processors cluster, and showed that there exists a perforance gap between the EC2 provisioned cluster and local traditional scientific cluster. Wang et al. [13] presented a easureent study on the ipact of virtualization on Aazon EC2 platfor. Their findings indicated that virtualization causes instability and variation to network throughput and packet delay. Li et al. [7] developed a perforance and cost coparator, i.e. CloudCp, to easure cloud services fro different cloud providers. Their study deonstrated that there was no single winner who outperfored the other counterparts in all aspects of its cloud service offerings. Cooper et al. [2] developed Yahoo! Cloud Serving Benchark (YCSB) fraework to facilitate perforance coparison. Barker et al. [1] analyzed the ipact of virtualization on the perforance of latency sensitive applications in the cloud. Furtherore, in exploiting heterogeneity in the cloud, there exist several studies. Suneja et al. [10] proposed to use Graphics Processing Uint (GPU) acceleration to speed up cloud anageent tasks in Virtual Machine Monitor (VMM). Lee et al. [6] introduced a scheduling echanis in the cloud that takes into consideration heterogeneity of the underlying platfor and workloads. Through atheatical odeling, Yeo et al. [14] found out that in order to achieve optial perforance, the perforance variation aong a heterogeneous cloud infrastructure should be no larger than three ties. To the best of our knowledge, there is no work focusing on exploiting the heterogeneity within the sae instance type of Aazon EC2, which otivates our work in this paper. 3 Micro-benchark In this section, we first analyze the hardware configuration of Aazon EC2. Then we utilize several icrobenchark tools to evaluate the perforance of various sub-types of instances. Specifically, standard large instance (1.large) is selected as the representative for perforance evaluation. 3.1 Hardware Configurations of EC2 We acquire the hardware inforation of Aazon EC2 instances by using cpuid coand, a non-trapping instruction that can be used in user ode without triggering trap to the underlying processor. Thus, the hypervisor does not capture the instruction and return odified results. Furtherore, we run cat /proc/cpuinfo coand to verify the results fro cpuid. The CPU odels fro both sources are identical, and the results are listed Table 1: Hardware configuration Instance type CPU odel %(2011) %(2012) E % 12% 1.sall E % 38% E5645 3% 30% 2218HE 18% 20% E % 40% E % 17% 1.large E5645 5% 42% 2218HE 4% 1% 270 4% - E % 6% 1.xlarge E % 46% E % 48% 270 2% - in Table 1. It is noteworthy that we only list the standard instance faily in Table 1. Diversified hardware is also used in high-cpu instance faily (c1.ediu and c1.xlarge). We exclude the due to space liit. Furtherore, the high-eory instances use identical Intel X5550 processors, and the cluster copute and cluster GPU instances both use Intel Xeon X5570 processors. We collected hardware inforation within two periods of tie to investigate the hardware changes fro hardware upgrade or replaceent. One period is fro April through July in 2011; the other one is fro January through March in For each period, we collect hardware inforation of 200 instances for each instance type, covering all availability zones in the US (Virginia) east region. The percentage of each CPU odel is shown in %(2011) and %(2012) coluns, respectively. The 2218HE and 270 odels are fro AMD Opteron series, whilst the rest are fro Intel Xeon series. Fro Table 1, it is clearly shown that newer processor odels are replacing older ones gradually, whilst the older ones are likely used for saller instances in the sae instance faily. For exaple, in 1.large instance, the AMD Opteron 270 (released in 2005) processor that was found in 2011 is no longer accessible in 2012, whilst the Intel Xeon E5645 (released Q1 10) CPU odel is ore frequently accessible in 2012 than in This trend is siilar in all standard (including 1.sall, 1.large, and 1.xlarge) and high-cpu (including c1.ediu and c1.xlarge) instances. Furtherore, we notice that the probability of a specific type of processor, e.g. E5645, significantly varies in different availability zones. In one availability zone, we can acquire 95% of instances hosted by E5645 achines, whilst in another zone, the probability of E5645 instances is as low as 10%. We conjecture that the availability zone with 95% of E5645 achines is a newly built 2

3 Score Instance E E E E E E Figure 1: UnixBench score, one and two processes Requests per second 6 x E5645 E5430 E Nuber of clients Figure 2: Redis SET operation data center within the US east region. The interesting question to ask is whether the heterogeneous hardware configuration within the sae instance type leads to diversified perforance. We select the 1.large instance as the exaple to evaluate perforance because this instance has a relatively large aount of eory and can be used in various general applications. 3.2 Micro-bencharks We use several icro-benchark tools to easure the perforance of 1.large instance, including UnixBench [11] to easure the CPU, Redis [9] to easure the eory, and Dbench [3] to easure the disk subsystes. To provide apples-to-apples coparison, we use the sae Aazon Machine Iage with CentOS5.6 in all the instances tested. The benchark is the only process running when we conduct the easureents. CPU perforance: UnixBench [11] utilizes ultiple tests to easure various aspects of the syste s perforance, priarily CPU s perforance. The test results are copared to the baseline syste to produce an index value. The entire set of index values are then cobined to ake a coposite index for the syste. To easure the likely diversity of instances fro the sae hardware configuration, we choose 20 instances fro each subtype of instance, i.e. E5507, E5430, and E5645. The results of the UnixBench benchark are shown in Fig. 1. The figure clearly deonstrates that the differences aongst the sae sub-type of instances, e.g. E5507, is sall, whilst the differences between different sub-types are significant. If one process is running, E5430 and E5645 are coparable in ters of perforance, whilst they are approxiately 1.15 ties of the perforance of E5507. When two processes are running, E5645 outperfors E5430, whilst E5430 further outperfors E5507. The perforance variation in ties is 1.21, and 1.1 ties for E5645, and E5430, respectively, wherein E5507 is taken as the baseline. Meory perforance: Redis [9] is an in-eory key-value store that has the benchark utility to siulate ultiple concurrent clients to send requests (e.g. SET, and GET) at the sae tie. In our easureents, we perfor 100,000 requests and vary the nuber of concurrent clients. Rando key is used to perfor the operations. The detailed results fro GET operations are depicted in Fig. 2. The results fro other operation are siilar to GET operation. Siilar to Fig. 1, in eory operations, E5645 instances outperfor E5430 and E5507 instances. The eory perforance of E5645 is 1.5 ties of that of E5507, whilst E5430 is 1.14 ties of E5507. Disk perforance: The results fro Dbench [3] show siilar trends as the Unixbench, and Redis. E5645 instances can provide disk throughput 1.25 ties as high as E5507 instances, whilst E5430 provides coparable disk throughput as E Application-level Benchark We use Httperf [5] to easure the Web server throughput. Dynaic HTTP request is used to ake the processor busy. Dynaic request eans after receiving a request fro a client, the Web server perfors a atheatical suation fro 1 through 100, and then returns the result to the client. Thus, the dynaic Web test is CPU-bound rather than network-bound. To try to avoid potential bottleneck fro client achine, we use a high- CPU ediu instance fro the sae zone acting as the client. The Httperf throughput results are depicted in Fig. 3. The figure deonstrates that the advantages fro separate subsystes, e.g. CPU, eory and disk, are accuulated at application-level, where E5645 is 1.6 ties as efficient as E5507 and E5430 is 1.2 ties as E

4 Response/sec Notation f h n p i x i C E5645 E5430 E Request/sec 4 Cost Analysis Figure 3: Httperf perforance Table 2: Notations Definition Hourly cost of an instance Nuber of hours to run Nuber of different instances Nuber of instances needed with worst perforance Probability of instances hosted with a specific hardware Perforance variation copared to the baseline instance The total cost Now we are aware that there exists various hardware configuration in the sae instance type. We analyze the potential cost saving by seeking for the best-perforing instances in the sae instance type. The worst-perforing instance is used as the baseline, the other instances are x (no less than 1) ties as fast as the baseline instance. We use the notations defined in Table 2. Given the sae aount of task (coputation, counication etc), with better-perforing instances, the task can be copleted with two alternatives: (1) saller nuber of instances running for the sae aount of tie; (2) sae nuber of instances running for shorter period of tie. Fro the cost perspective, these two alternatives are the sae. We take the first alternative as the exaple. The expected value of the perforance of a rando instance is defined as follows: E(X) = i=1 x i p i (1) The total cost of copleting the task, equivalent to n h hours work, using rando instances can be deduced as follows: C rando = n h f /E(X) (2) If we ai to select the best-perforing instances to coplete the task, the cost of this optiized scenario is: C opt = n h f /x opt (3) Furtherore, the trial and error testing process results in extra cost for the optiized scenario. As in Aazon EC2, the less than one hour usage is rounded up to and charged as one hour. Thus, the extra cost of finding n best-perforing instances is: C extra = n f /p opt (4) Here we assue that the test of finding one fast instance takes no ore than one hour and the jobs are relatively sall to the population of available servers. As a atter of fact, we can siply request for one instance fro Aazon, then inspect its cpuid. If the instance is not the best-perforing one, we siply discard it and request for another one. The potential cost saving is: C saving = C rando C opt C extra (5) Put Eq. 1, Eq. 2, Eq. 3, and Eq. 4 in Eq. 5, we can deduce the following equation: C saving = (h/( i=1 x i p i ) h/x opt 1/p opt ) n f (6) Understandably, if one fast instance is able to acquire cost gain, the total cost gain achievable fro ultiple instances grows linearly with the nuber of instances. This is also applicable to the price of the instance. Again, take the 1.large instance as the exaple. There are three different sub-types of instances, E5430, E5507, and E5645. The probability of each subtype of instance is 17%, 40%, and 42%, respectively. The unit cost of a regular 1.large instance (excluding reserved instances and spot instances) is $0.34/hour. The worstperforing instance is E5507, thus it is taken as the baseline. On average, E5430 and E5645 is 1.1 and 1.4 ties, respectively, as fast as E5507. Put all these values in Eq. 6, we can acquire the following equation: C saving = 0.34 n ( h 2.38) (7) In order to achieve cost saving, the requireent is C saving > 0, then we can get the necessity: h > That is to say, given the aforeentioned probability of each subtype of instance and its respective perforance, it starts to ake sense fro cost perspective to select E5645 instances to coplete the task if the required tie is larger than 18 hours. 4

5 Cost saving(%) p=0.1 p=0.5 p= Perforance variation (ties) Figure 4: Cost saving analysis Through two periods of several-onth easureents in 2011 and 2012, we found out that Aazon EC2 uses diversified hardware to host the sae type of instance. The hardware diversity results in perforance variation. In general, the variation between the fast instances and slow instances can reach 40%. In soe applications, the variation can even approach up to 60%. By selecting fast instances within the sae instance type, Aazon EC2 users can acquire up to 30% of cost saving, if the fast instances have a relatively low probability. In the future, we plan to investigate the scheduling echanis and analyze its ipact on the perforance of Aazon EC2 instances. If we have a task requires 100 E5507 coparable 1.large instances to coplete in a year (24hours/day*365days/year=8760 hours), the potential cost saving for the whole year is $40664, a 16% cost saving in percentage. Recall fro section 3.1 that different hardware is not distributed uniforly aong all the availability zones, but rather in soe zone one type of hardware doinates the whole zone, whist in another zone, another type of hardware doinates. Thus, it would also be interesting to analyze two types of hardware (e.g. E5507 and E5645) and investigate the axiu cost saving achievable. The result is depicted in Fig. 4, wherein p stands for the probability of the fast instances (e.g. E5645), and x-axis stands for the perforance variation in ties. Understandably, if the fast instances account for the ajority of the overall instances, e.g. p = 0.9, without a selection process, the probability of acquiring a fast instance is very high. Thus, the perforance is close to the optial situation with the selection process, and the cost saving achievable is trivial. However, as the fast instances account for less proportion of the overall instances, the cost saving achievable is becoing significant. In the case of p = 0.1, if the fast instance is 10 ties as fast as the slow instance, the cost saving is as high as 80%. Obviously, this is an unrealistic situation with all the efforts Aazon contributes to ake the sae type of instances function closely. Fro section 3.2 and 3.3, we know that ties variation is highly possible. With 1.5 ties variation, the achievable cost saving can reach 30%. For SMEs, which are the ajor custoers of Aazon EC2 platfor, this saving has a big ipact. 5 Conclusions In this paper, we investigated the hardware heterogeneity within the sae instance type of Aazon EC2. Standard large instance (1.large) was taken as the exaple. 6 Acknowledgents The research conducted in this paper has been funded by the Finnish funding agency for technology and innovation (Tekes) in Massive Scale Machine-to-Machine Service (MAMMotH) project (Dnro 820/31/2011). References [1] BARKER, S., AND SHENOY, P. Epirical evaluation of latencysensitive application perforance in the cloud. Proceedings of MMSys (2010), [2] COOPER, B., SILBERSTEIN, A., TAM, E., RAMAKRISHNAN, R., AND SEARS, R. Bencharking cloud serving systes with YCSB. Proceedings of SoCC (2010), [3] Dbench. [4] GARFINKEL, S. L. An evaluation of Aazon s grid coputing services: EC2, S3 and SQS. Tech. Rep. tr-08-07, Harvard University, [5] Httperf. httperf/. [6] LEE, G., CHUN, B., AND KATZ, R. H. Heterogeneity-aware resource allocation and scheduling in the cloud. Proceedings of HotCloud (2011), 1 5. [7] LI, A., YANG, X., KANDULA, S., AND ZHANG, M. CloudCp: coparing public cloud providers. Proceedings of IMC (2010), [8] PALANKAR, M., IAMNITCHI, A., RIPEANU, M., AND GARFINKEL, S. Aazon S3 for science grids: a viable solution? Proceedings of the 2008 international workshop on Data-aware distributed coputing (2008), [9] Redis. [10] SUNEJA, S., BARON, E., AND E. DE LARA, R. J. Accelerating the cloud with heterogeneous coputing. Proceedings of Hot- Cloud (2011), 1 5. [11] Unixbench. [12] WALKER, E. Bencharking aazon EC2 for high-perforance scientific coputing. USENIX ;login: 33, 5 (2008), [13] WANG, G., AND NG, T. The ipact of virtualization on network perforance of aazon ec2 data center. Proceedings of INFO- COM (2010), 1 9. [14] YEO, S., AND LEE, H. Using atheatical odeling in provisioning a heterogeneous cloud coputing environent. Coputer 44, 8 (2011),